An immune agent for web-based AI course.

To overcome weakness and faults of a web-based e-learning course such as Artificial Intelligence (AI), an immune agent was proposed, simulating a natural immune mechanism against a virus. The immune agent was built on the multi-dimension education agent model and immune algorithm. The web-based AI course was comprised of many files, such as HTML pages. JSPs, Java applets, and JavaScript programs. The immune agent was used to detect, recognize, eliminate, and repair the virus and faults of the AI course. The model, characteristics and algorithm of the immune agent were proposed and analyzed. The simulation example of the web-based AI course shows that the immune agent can detect some viruses and faults, recognize and eliminate the viruses and faults, and repair the e-learning system, keeping the e-learning system normal and robust.

**********

E-learning is becoming an important education mechanism of today's society, and the web-based AI course is an example of e-learning developed by Central South University for the Ministry of Education in China (Gong & Cai, 2002a, 2002b; Cai & Xu, 1996). First, the e-learning mechanism can expand the educational scope without limitation of time and place. Second, e-learning courses need fewer teachers for the same amount of students, and so it can address the lack of teachers for too many students in China. Third, e-learning platforms provide much more information and chances for students to develop their ability to think, create, and do it yourself (DIY).

However, sometimes viruses and faults occur in the e-learning course. First, a virus can destroy useful webpage files of the e-learning course. Second, a virus can copy itself to decrease the performance and efficiency of the e-learning platform, and rapidly spread viruses and faults all over the e-learning network. Third, a virus often causes a fault. Therefore, if a virus cannot be recognized and eliminated, and if the fault cannot be repaired, the e-learning course will be useless and dangerous. Hence, anti-virus and fault diagnosis are necessary and important for the e-learning courses to overcome their weakness and faults. The immune agent (IA) in this article is a good program, which has these functions.

Artificial immune system (AIS) is an adaptive system inspired from immunology and immune phenomena, to implement the immune model, principles, and functions through humans or computers (Cai & Gong, 2002). This technology has recently been advancing rapidly, and has shown the strong robustness for processing information and powerful ability to solve complex problems. The research evolved from immunology in the middle 1980s. In 1990, Bersini first used immune algorithms to solve problems in the adaptive network (Bersini & Varela. 1991). Meanwhile, Ishida (1990) used the artificial immune system to solve the fault diagnosis problem of the sensor network. In 1994, Forrest, Perelson, Allen, and Cherukuri first applied immune algorithms to computer security and virus detection (Dasgupta & Forrest, 1998). In 1995, Cooke and Hunt began to apply immune algorithms to the machine learning field. In 2002, Gong and Cai first proposed the multi-dimension immune network, used the immune mechanism to solve the security problem of the education network, and investigated the relations between robustness and immunity, then Gong (2003) investigated the multi-dimension education immune agent in his master thesis, and discussed how to use the immune computation techniques to build a new e-learning system. Recently, AIS is regarded as a new and good technique to solve web problems, like network security and fault diagnosis (Castro & Timmis, 2002).

IA PROBLEM AND MODEL

When a virus and fault occur in e-learning courses, e-learning systems should maintain complete and accurate data and knowledge, recognize and eliminate the virus, and repair the fault. Today's agents in the e-learning system are mostly passive to the virus and fault. And e-learning agents should be able to kill a virus, repair network faults, and provide special intelligent services autonomously.

In functionality, the e-learning agent has three dimensions, that is, teacher dimension, student dimension, and administrator dimension. So the e-learning agent model consists of four parts: (a) teacher module, (b) student module, (c) administrator module, and (d) common module (Gong & Cai, 2002a; Gong, 2003; Gong & Cai, 2003). The student dimension is the core of the e-learning agent, because students are the majority in the e-learning network. Only if students actively take full advantage of the e-learning resources, will the e-learning mechanism show usefulness and creativity. The common module refers to the common equipments of the agent model, like the communication interface, knowledge base, optimal algorithm, sensor, and so forth. The immune algorithm is just as important in the common module. The structure of the e-learning agent is shown in Figure 1, and the student module, the teacher module, and the administrator module interact with each other.

In Figure 1, the communication interface is an important part of the common module for the tri-dimension e-learning agent. Typing, text, symbolic language, and instructing language are valuable for the communication, which is composed of seven processes:

1. Intention: the teacher module decides what to teach students.

2. Generation: the teacher module decides what to talk to the student module with the standard instructing language.

3. Integration.

4. Perception: for sound media, the perception is phonic recognition; for print media, the perception is optical recognition.

[FIGURE 1 OMITTED]

5. Analysis: it consists of sentence understanding and semantic understanding.

6. Optimization: when a talk has more than one meaning, select one as the optimal solution to apply to the education agent.

7. Synthesis: the student module processes instructions with all kinds of factors.

The student module, the teacher module, and the administrator module belong to functional modules, and the three modules implement the student function, the teacher function, and the administrator function respectively. The common module, such as the immune computing module, belongs to the infrastructure module, and mainly provides the common equipment such as the computing environment, the communication interface, the human-machine interface, and so on.

The three function modules can be distributed and/or parallel. The tridimension e-learning architecture is a logic structure of the e-learning agent. For physical structure, most tasks of each dimension module can be done on the distributed server, such as the instructing agent server, the student agent server, the search agent server, and the management agent server, and so forth. So the task load can be distributed and balanced to overcome the load limitation. When the e-learning agent client accesses education services, the client interacts with the distributed servers through the tri-dimension interface: the teacher module interface, the student module interface, and the administrator module interface. Special service is done on the client of the e-learning agent. Public service will be done on the server of the e-learning agent and messages will be passed back to the client. Because all e-learning servers work independently, the distributed e-learning agent network can be parallel. Therefore, the whole network of the e-learning agent has higher efficiency. For example, the instructing agent server, the student agent server, the search agent server, and the managing agent server compute synchronously, and they can avoid affecting each other because their servers are independent and automatic.

Among the functional modules of the e-learning agent, the student module is the main functional module, for the e-learning educator advocates autonomous learning with the center of students, and the teacher module supports the student module, and the administrator module is used to serve the student module and the teacher module. However, for one kind of role the three modules are equal. For example, when a teacher accesses the e-learning agent, the user interacts with the teacher module of the e-learning agent, but the student module and the administrator module are invisible for the user. This is to say, a kind of user only can access the corresponding functional module, which is fit for the role of the user, while the other modules are all invisible for the user. This helps maintain security and the functional balance of the e-learning agent.

The module of immune computing is not only important to the individual e-learning agent, but also is necessary to keep e-learning web safe and robust, as shown in Figure 2.

In Figure 2, the e-learning web system is comprised of the tri-tier network architecture, the e-learning agents, and the AIS. The tri-tier network includes: the server tier, the database tier, and the application tier. AIS is comprised of the immune computing module, the core of immune computing, and the immune sensing layer.

The immune sensing layer is used to detect the virus and send the virus information to e-learning agents. Virus detection of the immune sensing layer is based on scanning the file system. The immune sensing layer detects the virus according to the following steps: (a) open the input source; (b) perform pattern matching in the virus-feature database; (c) if a match of virus is found, the sensing layer will send a virus alert to the immune computing core and e-learning agents; and (d) the sensing layer logs all events (Dasgupta & Forrest, 1998).

[FIGURE 2 OMITTED]

For the virus-feature database, its schema is comprised of tables, views, indexes, and so forth. For example, the feature table contains many fields, such as the ID number, the file name, and the feature. Initially, the content of the virus-feature database came from the real-world data profiles, such as the feature data of the love worm. For example, the virus-feature database of worms is shown in Table 1.

After the e-learning agent receives the virus alert, its immune computing module is activated. Then the e-learning agent will communicate with the core of immune computing. The e-learning agent has the following characteristics:

1. Immunity and robustness: there are often some viruses in the e-learning web, and the immune agent is required to resist the viruses. So robustness can be kept in the crucial parts of the e-learning web.

2. Virus detection: for a fault-tolerant architecture virus detection is a critical process. Most sophisticated recovery methods are only as good as the virus detection scheme that initiates their operations (Bradley & Tyrrell, 2002; Castro & Zuben, 2002). Virus detection is the process of scanning the file system and finding the malicious programs residing in the file system. File-based virus detection can be parallel.

3. Evolution: the e-learning web is like a sea of knowledge, the agent is a robot working in the sea, and the programs of the robot contain the intelligence of programmers. When human intelligence of the programmers evolves, the agent also maybe evolves.

4. Distribution and Parallelism. The e-learning web is a large-scaled network, and many agents are required to implement different missions. To increase efficiency, the agents can also be parallel, for example in grids.

5. Security: through the AIS, the e-learning web has better security. When the e-learning web is attacked, the e-learning agent activates its inner immune module, and recursively decreases the damage of the adventitious attack until the e-learning web recovers normal.

ALGORITHM DESIGN OF IA

The immune mechanism of the e-learning agent is inspired from natural immune system in human, animal etc (Dasgupta & Gonzalez, 2002; King, Russ, Lambert, & Reese, 2001; Toma, Endo, & Yamada, 2000). The natural immune system is a complex adaptive recognition system, and can protect the body against an adventitious virus. As a type of computation, the natural immune system is a parallel and distributed adaptive system.

The immune algorithm of the IA can memorize and learn the virus information (Ishida & Adachi, 1996; Dasgupta, 1998). The results of IA learning are all kinds of antibodies, which are used to recognize and eliminate viruses (Figure 3). When a virus invades the e-learning system, users may find some infected files, and then they can report the infected files to the core of immune computing and the e-learning agents.

The operations of the interface to activate the IAs through the user report include: (a) the users fill the report form by selecting many choice items; (b) the report form is transformed into some patterns of data, which can be understood by the core of immune computing; (c) the data are input into the core of immune computing; and (d) the IAs are activated by the messages from the core of immune computing.

[FIGURE 3 OMITTED]

On the other hand, the IAs can also scan the file system regularly and find the infected files. The operations include: (a) scan the target directories and read the feature information of every file in the directories; (b) match the feature information of the file in the virus-feature database; and (c) if a match is found, then the IAs are activated.

After the IA receives the information of the infected files, its immune algorithm will be activated. First, the IA initializes parameters and acquires the feature information of the virus from the infected files. The initialization procedure is used to give the initial values to some parameters and define some temporary arrays and variables. The feature information is read from the infected file by the file functions in the 10 package of Java 2. Second, the IA searches records in the database to match the feature information of the virus. If a record is matched, then the search process will be stopped and the corresponding method of eliminating the virus in the record will be called. Otherwise, when the search is done, no record is matched, the IA uses the reasoning rules to recognize and learn the feature information of the virus. The rule includes two parts: the first one is the condition of the virus feature and the second one is the conclusion of the rule-based reasoning, which shows the type of the virus and the elimination approach of the virus. For example, if the worm has the string "happy time" as a feature, then the worm is a kind of happy-time worm and the elimination approach of the virus is to delete the worm and its infected files. The rule matching is similar to the combination of DNA genes. Natural viruses are determined by some DNAs and computer viruses are determined by some reasoning rules. The immune algorithm is built on the random search of the rules. The reasoning rules are aggregated and classified from a lot of virus data by experts or machine learning. If the rules of the IA succeed to draw a conclusion, then the IA begins to eliminate the virus. Otherwise, other methods of fault diagnosis will be tried. Some standard manufacturing measures (e.g., the mean time between failures and availability) were used to analyze failure data on a mobile robot for reliability (Carlson & Murphy, 2003). A precision calibration procedure reduces the resulting errors by one order of magnitude (Chung, Ojeda, & Borenstein, 2001). Temporal fuzzy logic was used to represent monitoring knowledge and then utilized to effectively detect runtime failures (Lamine & Kabanza, 2000). The Bayesian view on uncertainty was used to solve fault detection and diagnosis problems in sampled-data stochastic systems (Berec, 1998). Multiple model estimation and neural network are used to predict the fault outcome, and to learn them collectively as a failure pattern (Goel, Dedeoglu, & Roumeliotis, 2000). The onboard MaKSI method for state estimation and fault diagnosis is particularly used in rovers (Washington, 2000). These methods for fault diagnosis can also be used in the hardware faults of the e-learning system in a similar way. After eliminating the virus, the useable files among the infected files need repairing. The approach of repairing the useable files is based on their backup files, and the process is to copy their backup files to their current directories.

IA EXAMPLE FOR Al WEB COURSE

In this article, the web-based AI course is a good test-bed of the IA example, in Figure 4.

The prototype development of the Al web course is a process of increasing developing from a small system. First, the knowledge points of the AI course are provided and the basic knowledge structure is established. Second, the themes, schedule, and difficulty of all web pages in the AI course are planned according to the instruction experiences and the e-learning model. The relations between one knowledge point and another, conjunctions among the themes, the interactivity between teaching and studying will be understood, and all the relations have to be represented with a formalized model, to embody better e-learning effects on computers. After the web-based AI course has been formalized with the model, the functions of the prototype for the e-learning system can be implemented with the advanced computing techniques such as the agent technique, Java, JSPs, XML, and so forth. Then, security, robustness, and intelligence of the e-learning system can be developed.

[FIGURE 4 OMITTED]

The definition of the IA example is created with Java. Here, suppose a mobile robot demo of the AI course was infected by the love worm, and the virus caused a fault, shown in Figure 5.

So the IA is activated to find the virus type from the virus-feature database. For the example, the virus is matched with the feature record of the love worm, shown in Figure 6.

And the virus and its infected files are eliminated immediately. But the useable files among the infected files are recorded for the further failover operation (failover means repairing the infected files, the fault files or the deleted files, in order to turn the E-learning system back to the normal state, i.e., no fault and no virus), shown in Figure 7.

After the IA recognizes and eliminates the virus, the dangerous source is eliminated and the web-based AI course is safe from the virus. After the IA repairs the deleted useable file among the defected files, the AI course recovers normal, shown in Figure 8.

From the IA example for the web-based AI course, it is certain that the IA can recognize and eliminate the virus, and repair the e-learning web. So the IA is able to keep the e-learning web immune and robust.

[FIGURE 5 OMITTED]

[FIGURE 6 OMITTED]

[FIGURE 7 OMITTED]

[FIGURE 8 OMITTED]

PERFORMANCE AND FUTURE WORK

The web-based AI course system has been tested and verified by the Ministry of Education from 2001 to 2002, and in the end the system has been certificated as excellent in China. The research of the immune agent has been continued in a master thesis (Gong, 2003), and in 2004 the thesis has been certificated excellent in the Hunan province of China. All in all, the system has been accepted and praised for its performance, rich contents, and intelligent functions.

However, more work will be done in the future, because the virus crisis is becoming a dangerous and crucial problem in the world (Balthrop, Forrest & Newman, 2004). As an effective and key anti-virus technique, the IA will recieve more attention. The two difficult issues for the IA will be the virus detection technique and the immune learning technique. On the other hand, the virus database will be built all over the world through the Internet. The grid computing technique will be used to provide a better platform for the immune computation and the IA.

CONCLUSIONS

It took about one and half years to developer the web-based AI course, and at last the course is assessed as excellent in China. Furthermore, the advanced research and development (R & D) has been done to increase its immunity, intelligence, security, and robustness. The good approach is just the IA technique, which is important and necessary for e-learning on the Web. In this article, the IA for the web-based AI course has been proposed. At first, the AI problem has been analyzed, and the AI model and characteristics have been proposed. Second, the immune algorithm of the IA has been proposed. At last, through the prototype implement of the IA in the web-based AI course, the IA and immune algorithm have been verified to overcome some viruses on the e-learning Web. Now we are making the IA and immune algorithm more powerful and intelligent for more viruses. In the future, the IA will surely become a required component of the e-learning web.

References

Balthrop, J., Forrest, S., Newman, M. E. J., & Williamson, M. M. (2004). Technological networks and the spread of computer viruses. Science,304(5670), 527-529.

Berec, L. (1998). Multi-model method to fault detection and diagnosis: Bayesian solution. An introductory treatise. International Journal of Adaptive Control and Signal Processing, 12(1), 81-92.

Bersini, H., & Varela, F. (1991), Hints for adaptive problem solving gleaned from immune network. In H.P. Schwefel & H. Muhlenbein (Eds.), Parallel problem solving from nature (pp. 343-354). London: Springer-Verlag.

Bradley, D. W., & Tyrrell, A. M. Immunotronics-novel finite-state-machine architectures with built-in self-test using self-nonself differentiation. IEEE Transactions on Evolutionary Computation, 6(3), 227-238.

Cai, Z. X., & Gong, T. (2002). Multi-dimension education immune network. Proceedings of the 6th World Multi-Conference on Systemics, Cybernetics and Informatics, I, 32-37.

Cai, Z. X., & Xu, G.Y. (1996). Artificial intelligence: Principles and applications (2nd ed.). Beijing: Tsinghua University.

Carlson, J., & Murphy, R R. (2003). Reliability analysis of mobile robots. Proceedings of IEEE International Conference on Robotics and Automation, 1, 274-281.

Castro, N. L. de, & Timmis, J. I. (2002). Artificial immune systems: A new computational intelligence approach. London: Springer-Verlag.

Castro, N. L., de, Zuben, J. R, von (2002). Learning and optimization using the clonal selection principle. IEEE Transactions on Evolutionary Computation, 6(3), 239-251.

Chung, H., Ojeda, L., & Borenstein, J. (2001). Sensor fusion for mobile robot dead-reckoning with a precision-calibrated fiber optic gyroscope. Proceedings of the 2001 IEEE International Conference on Robotics and Automation (pp. 3588-3593). Seoul, South Korea.

Cooke, D. E., & Hunt, J. E. (1995). Recognizing promoter sequences using an artificial immune system. Proceedings of the Intelligent Systems in Molecular Biology (pp. 89-97). Atlanta, GA: AAAI Press.

Dasgupta, D. (1998). Artificial immune system as a multi-agent decision support system. Proceedings of the IEEE International Conference on Systems, Man and Cybernetics, 4, 3816-3820.

Dasgupta, D., & Gonzalez, F. (2002). An immunity-based technique to characterize intrusions in computer networks. IEEE Transactions on Evolutionary Computation, 6(3), 281-291.

Forrest, S., Perelson, A.S., Allen, L., & Chdrukuri, R. (1994). Self-nonself discrimination in a computer. Proceedings of the 1994 IEEE Symposium on Research in Security and Privacy (pp. 202-212).

Goel, P., Dedeoglu, G., Roumeliotis, S. I., & Sukhatme, G. S. (2000). Fault detection and identification in a mobile robot using multiple model estimation and neural network. Proceedings of IEEE International Conference on Robotics and Automation, 3, 2302-2309.

Gong, T. (2003). Research on multi-dimension education immune agent. Unpublished master's thesis, Central South University, Changsha.

Gong, T, & Cai, Z. X. (2002a). Building and application of multi-dimension education agent. China Education info, 76, 55-56.

Gong, T, & Cai, Z. X. (2002b). Robust distributed system of multi-dimension education agents. Proceedings of International DCABES 2002, (pp. 144-150), Wuxi, China.

Gong, T., & Cai, Z. (2003). Immune mechanism & applications of education agent. Computer Engineering and Application, 39(24), 77-79.

Ishida, Y. (1990). Fully distributed diagnosis by PDP learning algorithm: Towards immune network PDP model. Proceedings of IJCNN 90, San Diego.

Ishida, Y.; & Adachi, N. (1996). Immune algorithm for multiagent: Application to adaptive noise neutralization. Proceedings of the IEEE International Conference on Intelligent Robots and Systems 1996, 3,1739-1746.

King, R. L, Russ, S. H., Lambert, A. B., & Reese, D. S. (2001). Artificial immune system model for intelligent agents. Future Generation Computer Systems, 17(4), 335-343.

Lamine, K., & Kabanza, F. (2000). History checking of temporal fuzzy logic formulas for monitoring behavior-based mobile robots. Proceedings of the 12th IEEE International Conference on Tools with Artificial Intelligence, (pp. 312-319).

Toma, N., Endo, S., & Yamada, K., & Miyagi, H. (2000). Immune distributed competitive problem solver with major histocompatibility complex and immune network. Proceedings of the IEEE International Conference on Systems, Man and Cybernetics (pp. 1865-1870).

Washington, R. (2000). On-board real-time state and fault identification for rovers. Proceedings of IEEE International Conference on Robotics and Automation, 2,1175-1181.

Note

I thank support of the NSFC project (60404021, 60234030) and the excellent doctoral degree project of Central South University (040125).

TAO GONG AND ZIXING CAI

Central South University, Changsha, Hunan China

[e-mail]

Table 1 Virus-Feature Database Example

No.  Type  Name        feature     language

1    worm  love        copy        VBScript
2    worm  Happy time  happy time  VBScript